app.py

import os
import spaces
import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import HfApi
import time

HF_TOKEN = os.environ.get("HF_TOKEN")
MODELS = ["Qwen/Qwen2.5-Coder-0.5B", "Qwen/Qwen2.5-Coder-1.5B"]
device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = MODELS[0]
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

def load_model(repo_id: str, progress = gr.Progress(track_tqdm=True)):
    global model, tokenizer
    api = HfApi(token=HF_TOKEN)
    if not api.repo_exists(repo_id=repo_id, token=HF_TOKEN): raise gr.Error(f"Model not found: {repo_id}")
    model = AutoModelForCausalLM.from_pretrained(repo_id, torch_dtype="auto", device_map="auto")
    tokenizer = AutoTokenizer.from_pretrained(repo_id)
    gr.Info(f"Model loaded {repo_id}")
    return repo_id

@spaces.GPU(duration=30)
def infer(message: str, sysprompt: str, tokens: int=30):
    messages = [
        {"role": "system", "content": sysprompt},
        {"role": "user", "content": message}
    ]
    
    input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text=[input_text], return_tensors="pt").to(model.device)
    start = time.time()
    generated_ids = model.generate(**inputs, max_new_tokens=tokens)
    end = time.time()
    elapsed_sec = end - start
    generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(inputs.input_ids, generated_ids)]
    output_str = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

    print(f"Input: {message}")
    print(f"Output: {output_str}")
    print(f"Elapsed time: {elapsed_sec} sec.")

    output_md = f"### {output_str}"
    info_md = f"### Elapsed time: {elapsed_sec} sec."

    return output_md, info_md

with gr.Blocks() as demo:
    model_name = gr.Dropdown(label="Model", choices=MODELS, value=MODELS[0], allow_custom_value=True)
    with gr.Row():
        message = gr.Textbox(label="Message", value="", lines=1)
        sysprompt = gr.Textbox(label="System prompt", value="You are Qwen, created by Alibaba Cloud. You are a helpful assistant.", lines=4)
        tokens = gr.Slider(label="Max tokens", value=30, minimum=1, maximum=2048, step=1)
    run_button = gr.Button("Run", variant="primary")
    output_md = gr.Markdown("<br><br>")
    info_md = gr.Markdown("<br><br>")

    gr.on(triggers=[run_button.click, message.submit], fn=infer, inputs=[message, sysprompt, tokens], outputs=[output_md, info_md])
    #run_button.click(infer, [message, sysprompt, tokens], [output_md, info_md])
    model_name.change(load_model, [model_name], [model_name])

demo.launch()